SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning
Kaushik Roy, Giovanni Salvatore D'urso, Nicholas Lawrance, Brendan Tidd, Peyman Moghadam
AI summary
Problem
Lifelong imitation learning struggles with catastrophic forgetting when agents sequentially acquire new skills, as existing distillation methods rely on noisy L2 feature matching that fails to preserve the intrinsic geometric structure of task representations.
Approach
The authors propose SPREAD, a framework that uses singular value decomposition to align the dominant low-rank subspaces of multimodal features across sequential tasks, combined with a confidence-guided KL divergence loss on high-probability action samples to stabilize policy transfer.
Key results
- State-of-the-art success rates across LIBERO-OBJECT, -GOAL, and -SPATIAL suites
- Significantly lower Negative Backward Transfer indicating reduced catastrophic forgetting
- Preservation of intrinsic low-dimensional feature manifolds across sequential tasks
- Theoretical proof that subspace alignment outperforms raw feature matching
Why it matters
It provides a robust, geometry-aware distillation strategy that enables scalable and reliable lifelong learning for robotic agents operating in dynamic, open-world environments.
Abstract
A key challenge in lifelong imitation learning (LIL) is enabling agents to acquire new skills from expert demonstrations while retaining prior knowledge. This re- quires preserving the low-dimensional manifolds and geometric structures that underlie task representations across sequential learning. Existing distillation methods, which rely on L2-norm feature matching in raw feature space, are sensitive to noise and high-dimensional variability, often failing to preserve intrinsic task manifolds. To address this, we introduce SPREAD, a geometry-preserving framework that employs singular value decomposition (SVD) to align policy representations across tasks within low-rank subspaces. This alignment maintains the underlying geometry of multimodal features, facilitating stable transfer, robustness, and generalization. Additionally, we propose a confidence-guided distillation strategy that applies a Kullback–Leibler divergence loss restricted to the top-M most confident action samples, emphasizing reliable modes and improving optimization stability. Experiments on the LIBERO, lifelong imitation learning benchmark, show that SPREAD substantially improves knowledge transfer, mitigates catas- trophic forgetting, and achieves state-of-the-art performance.